Claude Sonnet 5.5: API Specs, Effort, and Costs

Anthropic announced Claude Sonnet 5.5 on September 28, 2026 as an upgrade for everyday coding and professional work. The model sits alongside Opus 5.5 in the Claude family. Its practical appeal is a combination of application-building capability, clearer communication, and a lower token rate than the flagship tier.

These specifications describe the OpenRouter route checked on October 3. The listed inputs include text, images, and files. OpenRouter model page

Anthropic reports 70.6% on Terminal-Bench 4.0, output generation more than 30% faster than Sonnet 5, and costs up to 30% lower per task in its tests despite unchanged base token pricing. It also notes that Claude apps default to Medium effort while the Claude Platform defaults to High. These are vendor findings, not results reproduced by this article. Anthropic’s announcement

The distinction between token price and task cost is useful when choosing an agent. A workflow that batches useful tool operations or needs fewer repairs may finish with a smaller bill at the same price per million tokens. An evaluation should include failed calls and retries, rather than count only the tokens in the answer that was eventually accepted.

Define a representative set of feature-building, debugging, and document tasks, then compare effort settings on that fixed set. For code, include a test that can falsify a plausible-looking solution. For slides or spreadsheets, verify the exported artifact and its calculations. A visual preview alone does not confirm that a download or formula behaves correctly.

A hypothetical request with 100,000 input tokens and 10,000 output tokens costs $0.30 at the listed base rates, before caching and additional operations. That number is a token arithmetic example, not an observed coding session.

Our Sonnet 5.5 HTML-artifact gallery preserves an earlier session’s crane, globe, and wizard-game files. That article states the extent of its interaction checks. Sonnet 5.5 was paid in the newest-model catalog, so this pass did not run another three-prompt benchmark or reinterpret the earlier gallery as a controlled comparison.